Companies will shift the bulk of artificial intelligence spending from training to production inference in 2026, with inference accounting for roughly two-thirds of AI compute expenditures by the end of the year, according to InfoQ's 2026 AI, ML and Data Engineering Trends Report released Oct. 30. The report charts technology adoption from innovators to late majority and identifies small language models, reasoning systems, agent orchestration, inference engines and physical AI as the leading forces shaping the year ahead.
The report updates InfoQ's annual adoption graph using Geoffrey Moore's Crossing the Chasm framework, tracking how technologies move from early innovators to mainstream use. This year's edition adds eight new technologies to the innovators category: agent-to-agent protocols, inference engines, harness engineering, spec-driven development, AI software factories, physical AI, brain-computer interfaces and specialized decision models. Language models have pivoted from large-scale architectures to smaller models, while reasoning models now process math and logic before generating responses. AI agents have evolved from personal assistants to team-level tools and now to organization-wide systems. The report promotes 11 technologies to early majority, including context engineering, generative AI, retrieval-augmented generation, knowledge graphs, MLOps, vector databases, data contracts, data observability, graph analytics, synthetic data generation and open table formats. Eight technologies advanced to late majority: automated machine learning, image recognition, digital assistants, data mesh, stream processing, recommendation engines, lakehouses and Spark. The report dropped five mature topics from tracking entirely—natural language processing, deep learning, NoSQL databases, in-memory data grids and Hadoop.
Daniel Dominguez, one of the report's contributors, said AI is transitioning from an interface to an execution layer where "agents plan, use tools, execute actions and iterate until tasks complete." Hien Luu noted that stacks like vLLM, SGLang and llm-d simplify hosting open-weight models internally, which he said matters for financial firms handling private data and teams looking to reduce coding expenses. The report declares 2026 "the year of evaluations," with Luu emphasizing that assessments must examine the complete agent workflow rather than just model output. Metrics now extend beyond accuracy to include task completion, cost, time, interventions and safety, according to the report. Dominguez added that once an agent takes actions, evaluating the model alone becomes inadequate.
The surge in inference spending stems from reasoning models, extended context windows and agents transforming inference into a complex systems challenge that spans compute, memory, networking, routing and economics, the report explains. Physical AI—which links software intelligence to robots, drones, autonomous vehicles and industrial systems—introduces new risks because errors in physical systems carry real-world consequences, making reliability, simulation, safety and evaluation crucial. Chip Huyen noted that robotics lacks the ability to backtrack: a robot that falls breaks, and the state can't be reset. Harness engineering has expanded agent software development kits from basic loops to memory management, co-execution and file-system access. Luu said coding agents are powerful but require planning first: build a specification, brainstorm with the agent, then produce code. Agent-to-agent protocols such as MCP and A2A are emerging as an enterprise integration layer, while System One models like TypeSafe AI's Jev deliver deterministic inference with faster response times and lower costs than standard large language models.
InfoQ predicts increased US robotics innovation, cheaper and faster models, broader adoption of Jev-style decision models, and agent-to-agent communication becoming standard practice. Energy, chip supply, infrastructure and geopolitics will dominate 2027 coverage beyond the models themselves. The report urges practitioners to automate daily workflows rather than fear the transition. The shift toward inference-heavy workloads and physical AI deployments will pressure enterprises to rethink infrastructure architecture at a time when many organizations are still adjusting procurement cycles built around training-first assumptions. Companies betting on narrow orchestration strategies may find themselves constrained as cross-system agent coordination becomes table stakes for competitive advantage.